Tabular Transformers typically tokenize each feature as a token, so sequence length grows with the number of columns and attention cost scales as O(F2). We introduce EvoPatch, an offline tokenizer that learns a patch tokenization: it groups columns into K patches and represents each patch as a pooled token. EvoPatch builds a leakage-safe feature space from Gauss-rank normalized numeric features and leave-one-out target-encoded categorical features, constructs a feature similarity graph, and searches for a partition with multi-objective evolution. The search maximizes a fast validation proxy while maximizing token reduction, with a memetic refinement step that balances information across patches for stable solutions. The selected partition is computed once per dataset and reused unchanged for both training and inference, reducing attention cost without additional test-time computation and improving the computational efficiency of attention models on wide tables. Used as a tokenizer, EvoPatch compresses inputs for FT-Transformer, TabTransformer, and AutoInt on seven public datasets, achieving 17.5–75.4% token reduction with the largest absolute relative change in accuracy/RMSE bounded by 3.1%, and yields up to 2.8× training and 3.0× inference speedup on the widest dataset.

We propose a novel diffusion-based framework for automatic colorization of Anime-style facial sketches, which preserves the structural fidelity of the input sketch while effectively transferring stylistic attributes from a reference image. Our approach builds upon recent continuous-time diffusion models, but departs from traditional methods that rely on predefined noise schedules, which often fail to maintain perceptual consistency across the generative trajectory. To address this, we introduce SSIMBaD (Sigma Scaling with SSIM-Guided Balanced Diffusion), a sigma-space transformation that ensures linear alignment of perceptual degradation, as measured by structural similarity. This perceptual scaling enforces uniform visual difficulty across timesteps, enabling more balanced and faithful reconstructions. On a large-scale Anime face dataset, SSIMBaD attains state-of-the-art structural fidelity and strong perceptual quality, with robust generalization to diverse styles and structural variations.